The agent that grows with you
Python aiai-agentai-agentsanthropicchatgptclaudeclaude-codecodexhermeshermes-agentllmnous-researchopenai
"""Post-loop turn finalization for ``run_conversation``.
Budget summary, trajectory save, persist, diagnostics, response transforms, result
assembly, steer drain, memory/skill review. Synchronous, single return. ``logger`` is
imported lazily from ``agent.conversation_loop`` (no cycle, same logger name)."""
from __future__ import annotations
import logging
import os
import time
from contextlib import suppress
from typing import Any, Callable, List, Optional, Tuple
from agent.codex_responses_adapter import _summarize_user_message_for_log
from agent.delegation_context import is_dispatcher_owned_worker_context
from agent.interrupt_control import interrupted_during_api_call_reason
from agent.turn_failure_copy import exit_reason_failure, stamp_failure
from agent.context_compressor import _DB_PERSISTED_MARKER
from agent.message_content import flatten_message_text
from agent.message_metadata import append_message, stamp_message_timestamp
from agent.message_sanitization import _sanitize_surrogates
from agent.served_model import result_model_fields
# Verification-continuation nudges (verify-on-stop / pre_verify) must be stripped from
# returned/live history to avoid role-alternation breaks; the assistant response is
# real content and is not flagged. (#65919)
_VERIFICATION_CONTINUATION_FLAGS = ("_verification_stop_synthetic", "_pre_verify_synthetic")
_SENTENCE_END = {".", "!", "?", "。", "!", "?", "`", ")"}
# ``result[key] = agent.session_<key>`` for the per-session usage/cost counters.
_SESSION_TOKEN_KEYS = (
"input_tokens", "output_tokens", "cache_read_tokens", "cache_write_tokens",
"reasoning_tokens", "prompt_tokens", "completion_tokens", "total_tokens",
)
_SESSION_COST_KEYS = ("estimated_cost_usd", "cost_status", "cost_source")
def _assistant_row_missing_visible_text(msg: dict) -> bool:
"""True when an assistant row has no visible text (blank final or tool-only)."""
if not isinstance(msg, dict) or msg.get("role") != "assistant":
return False
return not flatten_message_text(msg.get("content")).strip()
def _record_kanban_budget_exhausted(
kanban_task: str, api_call_count: int, max_iterations: int, logger: logging.Logger
) -> None:
"""Record a terminal ``timed_out`` outcome for a kanban worker out of budget.
Routed via ``_record_task_failure`` (not ``kanban_block``) so it counts toward the
consecutive-failure circuit breaker. Idempotent via the ``_end_run`` CAS
(``WHERE ended_at IS NULL``), so safe from multiple exit paths.
This is a bounded fallback (#87096): the CAS invariant in ``_end_run`` (``WHERE ended_at IS NULL``)
guarantees idempotence — if another path already closed the run this is a no-op — so it is safe to call
from multiple exit paths.
"""
try:
from hermes_cli import kanban_db as _kb
from hermes_cli import kanban_db_connect as _kbc
from hermes_cli import kanban_db_dispatch as _kbd
_conn = _kbc.connect()
try:
_kbd._record_task_failure(
_conn,
kanban_task,
error=(
f"Iteration budget exhausted ({api_call_count}/{max_iterations}) — "
"task could not complete within the allowed iterations"
),
outcome="timed_out",
release_claim=True,
end_run=True,
event_payload_extra={"budget_used": api_call_count, "budget_max": max_iterations},
)
finally:
with suppress(Exception):
_conn.close()
except Exception:
logger.warning(
"Failed to record budget-exhausted failure for task %s", kanban_task, exc_info=True
)
def _drop_verification_continuation_scaffolding(messages) -> None:
"""Remove verification-continuation nudges in place; only the synthetic nudges carry
these flags, so the real attempted final answer persisted to state.db survives."""
messages[:] = [
m for m in messages
if not (isinstance(m, dict) and any(m.get(f) for f in _VERIFICATION_CONTINUATION_FLAGS))
]
def _clone_background_review_messages(messages):
"""Copy the review input without aliasing the live transcript."""
# Lazy: conversation_loop imports this module (cycle).
from agent.conversation_loop import _clone_message_for_send
return [_clone_message_for_send(message) for message in messages]
def _invoke_hook_safely(name: str, logger: logging.Logger, **kwargs) -> list:
"""Fire a lifecycle plugin hook; a failing hook is logged, never fatal."""
try:
from hermes_cli.lifecycle import invoke_hook
return invoke_hook(name, **kwargs)
except Exception as exc:
logger.warning("%s hook failed: %s", name, exc)
return []
def _guarded_cleanup(label: str, fn: Callable[[], Any], errors: list[str], logger) -> None:
"""Post-loop cleanup must never lose the response: each step is guarded
independently and errors surface via ``cleanup_errors`` (#8049)."""
try:
fn()
except Exception as err:
errors.append(f"{label}: {err}")
logger.error("finalize_turn: _%s failed: %s", label, err, exc_info=True)
def _resolve_budget_fallback(
agent, *, final_response, api_call_count, interrupted, failed, messages, _turn_exit_reason,
_pending_verification_response, _pending_verification_response_previewed, logger,
) -> tuple[Any, Any, bool, Any]:
"""Iteration-budget exhaustion. Returns ``(final_response, _turn_exit_reason,
preserved_verification_fallback, interrupted)``."""
budget_exhausted = (
api_call_count >= agent.max_iterations or agent.iteration_budget.remaining <= 0
)
preserved_verification_fallback = False
if (
final_response is None and budget_exhausted and not interrupted and not failed
and str(_turn_exit_reason) in {"unknown", "budget_exhausted"}
):
_turn_exit_reason = f"max_iterations_reached({api_call_count}/{agent.max_iterations})"
if _pending_verification_response:
# A verification gate withheld a composed answer, then the budget ran out:
# preserve it rather than make another fallible call. The explicit pending
# value is the provenance guard; unrelated error exits never enter here.
# Previewed only if the reused candidate was actually streamed as interim.
final_response = _pending_verification_response
if _pending_verification_response_previewed:
agent._response_was_previewed = True
# Reuses the candidate the user already saw sealed as an interim (#130396).
agent._reused_response_text = final_response
preserved_verification_fallback = True
else:
# _handle_max_iterations makes one extra toolless request for a summary.
agent._emit_diagnostic_status(
f"⚠️ Iteration budget exhausted ({api_call_count}/{agent.max_iterations}) "
"— asking model to summarise"
)
if not agent.quiet_mode:
agent._safe_print(
f"\n⚠️ Iteration budget exhausted ({api_call_count}/{agent.max_iterations}) "
"— requesting summary...", diagnostic=True,
)
_summary_start = time.time()
try:
final_response = agent._handle_max_iterations(messages, api_call_count)
except InterruptedError:
# The turn ends interrupted, so the pending interrupt message is returned
# for requeue instead of being cleared behind a fallback summary. A redirect
# also ends it: the budget is spent, so there is no loop to restart into.
from agent.conversation_loop import INTERRUPT_WAITING_FOR_MODEL_PREFIX
interrupted = True
_turn_exit_reason = interrupted_during_api_call_reason(agent)
final_response = f"{INTERRUPT_WAITING_FOR_MODEL_PREFIX}{time.time() - _summary_start:.1f}s elapsed)."
# A kanban worker must record a terminal outcome whether or not a fallback path
# was eligible, so the dispatcher learns the worker could not complete. Only the
# dispatcher-owned worker owns the task: an in-process delegate_task child or cron run
# inherits ``HERMES_KANBAN_TASK`` via os.environ but exhausting ITS budget must not
# close the parent's run and release its claim (#112817).
_kanban_task = (
os.environ.get("HERMES_KANBAN_TASK")
if budget_exhausted and is_dispatcher_owned_worker_context() else None
)
# If running as a kanban worker, signal the dispatcher that the worker could not complete (rather than
# treating it as a protocol violation). This applies whether the user-facing fallback came from the
# summary call or an explicitly pending continuation; both exhausted the task budget and must advance
# the failure circuit. We route through ``_record_task_failure(outcome="timed_out")`` rather than
# ``kanban_block`` so this counts toward the dispatcher's consecutive-failure circuit breaker (#29747
# gap 2).
# Bounded fallback (#87096): budget was exhausted but none of the normal fallback paths were eligible
# (interrupted / failed / anomalous exit_reason). If running as a kanban worker we must still record a
# terminal outcome so the task does not remain in an ambiguous lifecycle state. The worker's run is
# closed via ``_record_task_failure`` (compare-and-swap receipt path) which is a no-op if another path
# closed it — the CAS invariant in ``_end_run`` (``WHERE ended_at IS NULL``) guarantees idempotence.
if _kanban_task:
_record_kanban_budget_exhausted(_kanban_task, api_call_count, agent.max_iterations, logger)
return final_response, _turn_exit_reason, preserved_verification_fallback, interrupted
def _rollback_interrupted_preflight_display(agent, interrupted) -> None:
"""Roll back the preflight-seeded display count only when an interrupt wins before
any provider response; compaction state (incl. ``-1``) stays with the real-usage
path. Type-pinned guards keep MagicMock/SimpleNamespace doubles inert."""
_preflight_snapshot = getattr(agent, "_turn_preflight_display_snapshot", None)
if (
interrupted is True
and isinstance(_preflight_snapshot, int)
and not isinstance(_preflight_snapshot, bool)
and getattr(agent, "_turn_received_provider_response", False) is not True
and getattr(agent, "context_compressor", None) is not None
):
_rollback_fn = getattr(
agent.context_compressor, "rollback_interrupted_preflight_display_tokens", None
)
if callable(_rollback_fn):
_rollback_fn(_preflight_snapshot)
def _drop_transcript_scaffolding(agent, messages) -> None:
"""Strip private retry scaffolding first, or a later "continue" replays
assistant("(empty)") / recovery nudges into the same empty-response loop. Only
the synthetic verification nudges go; the assistant candidate persists (#65919)."""
agent._drop_trailing_empty_response_scaffolding(messages)
_drop_verification_continuation_scaffolding(messages)
def _recover_final_from_stream(agent, final_response, interrupted, failed) -> tuple[Any, bool]:
"""An empty terminal completion is not authoritative when the stream already
delivered text; recover before persist so a blank tail isn't frozen (#95514).
Returns ``(final_response, recovered_from_stream)``. Called by the finalizer BEFORE
the fallible tail-shaping/persist steps so the recovered text is already bound when
one of them raises — a persist failure must not lose text the user already saw."""
if interrupted or failed:
return final_response, False
_streamed = getattr(agent, "_current_streamed_assistant_text", "") or ""
_streamed = _streamed.strip() if isinstance(_streamed, str) else ""
if not (flatten_message_text(final_response).strip() if final_response else "") and _streamed:
return _streamed, True
return final_response, False
def _close_transcript_tail(agent, messages, final_response, interrupted, _recovered_from_stream) -> None:
"""Shape the transcript tail before the durable snapshot (scaffolding already dropped
and ``final_response`` already stream-recovered by the caller)."""
# An interrupt can leave a tool result as the tail; close the sequence so strict
# providers don't see ``tool → user`` (placeholder: final_response is usually empty).
if interrupted:
from agent.message_sanitization import close_interrupted_tool_sequence
close_interrupted_tool_sequence(messages, final_response, stopped=True)
# Recovery ``break`` sites can return a final_response with no closing assistant
# row; enforce "delivered final_response ⇒ assistant row" here. Compare content,
# not role, so a matching verification candidate isn't dup'd.
if final_response and not interrupted:
# Some recovery/fallback paths return a real final_response without adding a closing assistant
# message to the transcript (e.g. the partial-stream and prior-turn-content recovery ``break`` sites
# in ``conversation_loop``). If persisted as-is, the durable session can end at a tool/user message
# even though the caller — and the gateway platform — already saw a completed assistant response.
# The next turn then replays a user-only backlog and the model re-answers every "unanswered"
# message. Close the durable turn at the source, at the single chokepoint every recovery ``break``
# flows through, so the invariant "delivered final_response ⇒ assistant row in transcript" holds
# regardless of which path produced it. (#43849 / #44100) Compare content (not just role) so a
# verification candidate that matches the final response is not duplicated at budget exhaustion.
# (#65919 §7)
_tail = messages[-1] if messages else None
if not isinstance(_tail, dict) or _tail.get("role") != "assistant":
append_message(messages, {"role": "assistant", "content": final_response})
elif (
_tail.get("content") != final_response
and _assistant_row_missing_visible_text(_tail)
and (_tail.get("tool_calls") or _recovered_from_stream)
):
# Pure tool-call turn or stream-recovered blank (#95514): fill the persisted
# blank row's content rather than append a second row.
_tail["content"] = final_response
stamp_message_timestamp(_tail)
_tail.pop(_DB_PERSISTED_MARKER, None)
agent._db_flush_scan_prefix = None
# Request is complete, so replace API-local voice/model/skill guidance with the
# clean user input before the durable snapshot (earlier flushes still needed them).
# Earlier turn-start flushes use the DB-only override because their messages are still needed for the
# API request; this finalizer runs after that request is complete (#48677 / #63766).
_apply_override = getattr(agent, "_apply_persist_user_message_override", None)
if callable(_apply_override):
_apply_override(messages)
def _store_rewrite_on_tail(agent, messages, raw, final) -> None:
"""A recovery exit whose text is already the unflushed tail row (the iteration-budget summary, a
stop gate's answer kept as the budget fallback) stores the rewrite the user gets, not the model
text (#44239). A held turn's hidden answer becomes the published reply; the model already saw
it, so ``api_content`` keeps its bytes on the wire."""
tail = messages[-1] if messages else None
if not (isinstance(tail, dict) and tail.get("role") == "assistant" and tail.get("content") == raw
and not tail.get(_DB_PERSISTED_MARKER)):
return
if tail.get("display_kind") == "hidden" and getattr(agent, "_hold_stream", False) is True:
del tail["display_kind"]
if final != raw:
tail.setdefault("api_content", raw)
tail["content"] = final
def _micro_compact_after_turn(agent, messages, final_response, logger, task_id) -> None:
"""Post-turn micro-compaction: absorb the oldest uncompacted exchange into the
rolling summary before persist, amortizing compression across turns."""
try:
_compressor = getattr(agent, "context_compressor", None)
# Strict `is True` + callable gates: plugin context engines and MagicMock
# compressors pass duck checks and would wipe the transcript. Never run while
# compression.checkpoint_required is armed (no checkpoint hook here), nor for
# persistence-isolated agents (background review fork): that burns an aux-LLM
# call on a throwaway transcript and could compact the CANONICAL session rows.
if (
_compressor
and getattr(_compressor, '_micro_compact_enabled', False) is True
and callable(getattr(_compressor, '_micro_compact', None))
and final_response
and getattr(agent, "compression_checkpoint_required", False) is not True
and not getattr(agent, "_persist_disabled", False)
):
_before = len(messages)
# A rotating compaction earlier in this turn may have moved the agent to a child session.
_compacted = _compressor._micro_compact(
messages, turn_session_id=getattr(agent, "_inflight_turn_session_id", None),
)
# Defrag rewrites the newest MICRO marker in place and pops _db_persisted;
# the compressor flags us to invalidate the flush-scan cursor, else the
# rewritten row is identity-skipped (stale).
if getattr(_compressor, "_flush_scan_cursor_invalidated", False):
_compressor._flush_scan_cursor_invalidated = False
agent._db_flush_scan_prefix = None
if isinstance(_compacted, list) and _compacted:
_spliced = _compacted is not messages # no-op and defrag passes return the input
messages[:] = _compacted
if _spliced:
# The splice summarized tool results away: a repeat read must serve them
# again, not an "unchanged" stub pointing at a body that is gone (#32106).
from agent.conversation_compression import _reset_read_dedup_caches
_reset_read_dedup_caches(task_id, session_id=agent.session_id or "")
if _before != len(messages):
logger.info("Micro-compaction: %d -> %d messages", _before, len(messages))
except Exception as _mc_err:
logger.info("Micro-compaction failed: %s", _mc_err)
def _log_turn_exit(agent, messages, final_response, api_call_count, _turn_exit_reason, interrupted, logger) -> None:
"""Always INFO so agent.log captures WHY every turn ended; WARNING when the last
message is a tool result (the "just stops" scenario)."""
_touch = getattr(agent, "_touch_activity", None)
if callable(_touch):
# Stamp the activity clock at the loop's end (#131740): the watchdog's stall
# surface then names the finalizer ("turn end logged") instead of the last
# API call, and a turn that wedges in the post-loop tail is measured — and
# aborted — from when the loop actually finished, not from the last provider
# response. Never raises into the finalizer.
try:
_touch("turn end logged")
except Exception:
logger.debug("turn-end activity stamp failed", exc_info=True)
_last_msg_role = messages[-1].get("role") if messages else None
_last_tool_name = None
if _last_msg_role == "tool":
# Walk back to the assistant message with the tool call.
for _m in reversed(messages):
if _m.get("role") == "assistant" and _m.get("tool_calls"):
_tcs = _m["tool_calls"]
if _tcs and isinstance(_tcs[0], dict):
_last_tool_name = _tcs[-1].get("function", {}).get("name")
break
_turn_tool_count = sum(
1 for m in messages
if isinstance(m, dict) and m.get("role") == "assistant" and m.get("tool_calls")
)
# Fork turns (background review, side questions) carry ``_turn_origin``; tagging the
# exit line keeps a fork's ``interrupted_during_api_call`` from reading as a killed
# foreground stream — the fork shares the parent's session_id and often its model (#118693).
_turn_origin = getattr(agent, "_turn_origin", None)
_diag_msg = (
"Turn ended: reason=%s model=%s api_calls=%d/%d budget=%d/%d "
"tool_turns=%d last_msg_role=%s response_len=%d session=%s"
+ (" origin=%s" if _turn_origin else "")
)
_diag_args = (
_turn_exit_reason, agent.model, api_call_count, agent.max_iterations,
agent.iteration_budget.used if agent.iteration_budget else 0,
agent.iteration_budget.max_total if agent.iteration_budget else 0,
_turn_tool_count, _last_msg_role, len(final_response) if final_response else 0,
agent.session_id or "none",
*((_turn_origin,) if _turn_origin else ()),
)
if _last_msg_role == "tool" and not interrupted:
logger.warning(
"Turn ended with pending tool result (agent may appear stuck). "
+ _diag_msg + " last_tool=%s",
*_diag_args, _last_tool_name,
)
else:
logger.info(_diag_msg, *_diag_args)
def _append_file_mutation_footer(agent, final_response, logger):
"""Append the verifier advisory when ``write_file`` / ``patch`` calls failed and were
never superseded by a successful write to the same path (surfaces over-claiming)."""
try:
# File-mutation verifier footer. This catches the specific case — reported by Ben Eng
# (#15524-adjacent) — where a model issues a batch of parallel patches, half of them fail with
# "Could not find old_string", and the model summarises the turn claiming every file was edited. The
# user then has to manually run ``git status`` to catch the lie. With this footer the truth is
# surfaced on every turn, so over-claiming is structurally impossible past the model. Gate: only
# applied when a real text response exists for this turn and the user didn't interrupt.
# Empty/interrupted turns already have other surface text that shouldn't be augmented.
_failed = getattr(agent, "_turn_failed_file_mutations", None) or {}
if _failed and agent._file_mutation_verifier_enabled():
_failed = agent._file_mutations_still_failed(_failed)
footer = agent._format_file_mutation_failure_footer(_failed)
if footer:
final_response = final_response.rstrip() + "\n\n" + footer
except Exception as _ver_err:
logger.debug("file-mutation verifier footer failed: %s", _ver_err)
return final_response
def _explain_abnormal_exit(agent, final_response, _turn_exit_reason, preserved_verification_fallback, logger):
"""Turn-completion explainer: on abnormal exits, surface one explanation from
``_turn_exit_reason``. Only acts when no usable reply exists (empty, "(empty)",
or a short unpunctuated fragment); ``text_response(...)`` exits stay silent."""
try:
if not agent._turn_completion_explainer_enabled():
return final_response
_stripped = (final_response or "").strip()
_is_empty_terminal = _stripped in ("", "(empty)")
# A short fragment not from a text_response exit and lacking sentence-ending
# punctuation is treated as a truncated partial (#34452).
_is_partial_fragment = (
not _is_empty_terminal
and not preserved_verification_fallback
and not str(_turn_exit_reason).startswith("text_response")
and _turn_exit_reason != "fallback_prior_turn_content"
and len(_stripped) <= 24
and _stripped[-1:] not in _SENTENCE_END
)
if _is_empty_terminal or _is_partial_fragment or str(_turn_exit_reason) == "partial_stream_recovery":
_explanation = agent._format_turn_completion_explanation(
_turn_exit_reason, getattr(agent, "_last_persistence_error_cause", None),
db_path=getattr(getattr(agent, "_session_db", None), "db_path", None),
model=str(getattr(agent, "model", "") or ""),
)
if _explanation:
# Replace the bare sentinel; keep a partial fragment and append why.
final_response = _explanation if _is_empty_terminal else _stripped + "\n\n" + _explanation
except Exception as _exp_err:
logger.debug("turn-completion explainer failed: %s", _exp_err)
return final_response
def _last_turn_reasoning(messages) -> Optional[Any]:
"""Reasoning from the CURRENT turn only: stop at this turn's user message (#17055),
but take the most recent non-empty reasoning since many providers emit it on the
tool-call step and leave the final step with reasoning=None."""
for msg in reversed(messages):
if msg.get("role") == "user":
return None # turn boundary — don't cross into prior turns
if msg.get("role") == "assistant" and msg.get("reasoning"):
return msg["reasoning"]
return None
def _apply_output_hooks(
agent, final_response, logger, *, platform, effective_task_id, turn_id, original_user_message,
messages,
) -> tuple[Any, bool, Optional[Any]]:
"""Resolve the turn's ``transform_llm_output`` outcome, then fire ``post_llm_call`` once per
turn after the tool loop. Returns ``(final_response, transformed, pre_transform_response)``.
The transform itself normally already ran before the assistant row was first persisted
(``apply_llm_output_transform`` from ``finish_text_response`` / ``_persist_step``); this
call returns that recorded outcome, and only fires the hook here when no earlier seam saw a
response (e.g. text that only appeared through ``_explain_abnormal_exit``)."""
final_response, transformed, pre_transform = apply_llm_output_transform(
agent, final_response, turn_id=turn_id, platform=platform, logger=logger,
)
# Detached forks are internal work and must not publish turns under the parent's session ID.
if not getattr(agent, "_persist_disabled", False):
# Leased on the turn thread so the capture holds this turn's context; None unless a
# loaded plugin holds llm.side_question. The response must survive a failed capture.
try:
from agent.side_question_lease import issue_side_question_token
side_question_token = issue_side_question_token(agent, messages, turn_id)
except Exception:
side_question_token = None
logger.warning("side question lease failed", exc_info=True)
_invoke_hook_safely(
"post_llm_call", logger,
session_id=agent.session_id,
task_id=effective_task_id,
turn_id=turn_id,
user_message=original_user_message,
assistant_response=final_response,
conversation_history=list(messages),
model=agent.model,
platform=platform,
side_question_token=side_question_token,
)
return final_response, transformed, pre_transform
def apply_llm_output_transform(
agent, final_response, *, turn_id, platform=None, logger=None,
) -> tuple[Any, bool, Optional[Any]]:
"""Fire ``transform_llm_output`` once per turn and return
``(final_response, transformed, pre_transform_response)``.
Called BEFORE the final assistant row is first persisted — from ``finish_text_response``
ahead of its durable flush, and from ``finalize_turn._persist_step`` ahead of the
recovery-path tail close — so the text the user sees is the text stored in SQLite/JSON and
replayed next turn (#44239). SQLite treats a non-blank assistant row as settled (a re-flush
adopts the stored content rather than overwriting it), so transforming after that first
write can never reach the durable store. Idempotent per ``turn_id``: later callers in the
same turn get the recorded outcome instead of a second hook firing. Only the current
turn's not-yet-written text is touched — earlier turns and the system prompt are never
rewritten (prompt-cache invariant)."""
if logger is None:
from agent.conversation_loop import logger
recorded = getattr(agent, "_llm_output_transform", None)
if isinstance(recorded, tuple) and len(recorded) == 3 and recorded[0] == turn_id:
_, transformed, pre_transform = recorded
return final_response, transformed, pre_transform
if not final_response:
return final_response, False, None
if platform is None:
platform = getattr(agent, "platform", None) or ""
transformed, pre_transform = False, None
payload = dict(response_text=final_response, session_id=agent.session_id or "", model=agent.model,
platform=platform, turn_id=turn_id) # turn_id: per-turn identity for the hook callback gate
held = getattr(agent, "_held_transforms", None) if getattr(agent, "_hold_stream", False) else None
if held:
rewrite = _run_held_transforms(held, payload, logger)
else:
# First hook to return a string wins; None/empty leaves the text unchanged.
rewrite = next((r for r in _invoke_hook_safely("transform_llm_output", logger, **payload)
if isinstance(r, str) and r), None)
if rewrite:
pre_transform, final_response, transformed = final_response, rewrite, True
agent._llm_output_transform = (turn_id, transformed, pre_transform)
return final_response, transformed, pre_transform
HELD_TRANSFORM_FAILED_REPLY = "This reply was withheld: the plugin that rewrites replies in this session failed."
def _run_held_transforms(callbacks, payload: dict, logger) -> Optional[str]:
"""A held turn runs the transforms captured at its start (``held_stream_transforms``), not the live
registry a mid-turn reload empties. When none rewrote and one raised or timed out, the user gets a
fixed notice, never the text the plugin had to see first."""
from hermes_cli.plugins import get_plugin_manager
try:
rewrite, failed = get_plugin_manager().run_held_transforms(callbacks, **payload)
except Exception: # fail closed: whatever broke dispatch, the original stays withheld
logger.warning("held transform_llm_output dispatch failed", exc_info=True)
rewrite, failed = None, True
return HELD_TRANSFORM_FAILED_REPLY if failed else rewrite
def settle_held_final(agent, result: Any) -> None:
"""A held turn's reply leaves only through its transform snapshot. ``finish_text_response``,
``finalize_turn``, the Codex runtime and the truncation ceiling run it before persisting; every
other exit (provider errors, retry exhaustion, interrupts, compression failures, ...) returns its
text through ``run_conversation``, which calls this, so no path hands the user text the plugin
did not see first, and a failed transform yields the withheld notice."""
if getattr(agent, "_hold_stream", False) is not True or not isinstance(result, dict):
return
text, turn_id = result.get("final_response"), getattr(agent, "_current_turn_id", "") or ""
recorded = getattr(agent, "_llm_output_transform", None)
if not isinstance(text, str) or not text or (isinstance(recorded, tuple) and recorded[:1] == (turn_id,)):
return
final, transformed, _ = apply_llm_output_transform(agent, text, turn_id=turn_id)
if result.get("error") == text: # keep the mirror an error card reads
result["error"] = final
result.update(final_response=final, response_transformed=transformed, pre_transform_response=None)
def public_pre_transform(agent, pre_transform: Optional[Any]) -> Optional[Any]:
"""``pre_transform_response`` for a turn result: a held turn's original never leaves the agent
(receipts, replay-gap recovery, post-stream suffix painting all see only the rewrite)."""
return None if getattr(agent, "_hold_stream", False) else pre_transform
def _mirror_turn_to_memory(
agent, *, original_user_message, final_response, interrupted, messages, review_memory, review_skills, vetoed,
) -> None:
"""External-memory sync (+ next prefetch), then the background memory/skill review. A prompt a
``pre_llm_call`` hook vetoed never leaves the host, so neither runs for that turn."""
if vetoed:
return
agent._sync_external_memory_for_turn(
original_user_message=original_user_message, final_response=final_response,
interrupted=interrupted, messages=messages,
)
# Background memory/skill review runs AFTER delivery so it never competes with the
# user's task. Suppressed by skip_background_review (e.g. cron): the fork costs
# ~30K tokens / event with no human-in-the-loop benefit. Best-effort; the review
# clones the snapshot structurally so its sanitizers can't reach the live transcript.
if (
final_response
and not interrupted
and not getattr(agent, "skip_background_review", False)
and (review_memory or review_skills)
):
with suppress(Exception):
agent._spawn_background_review(
messages_snapshot=list(messages), review_memory=review_memory,
review_skills=review_skills,
)
def finalize_turn(
agent, *, final_response, api_call_count, interrupted, failed, messages, conversation_history,
effective_task_id, turn_id, user_message, original_user_message, _should_review_memory,
_turn_exit_reason, _pending_verification_response=None,
_pending_verification_response_previewed=False,
):
"""Run the post-loop finalization and return the turn ``result`` dict."""
from agent.conversation_loop import logger
from agent.turn_context import PRE_LLM_BLOCK_EXIT_REASON
vetoed = _turn_exit_reason == PRE_LLM_BLOCK_EXIT_REASON # a pre_llm_call block: no model call at all
final_response, _turn_exit_reason, preserved_verification_fallback, interrupted = _resolve_budget_fallback(
agent, final_response=final_response, api_call_count=api_call_count,
interrupted=interrupted, failed=failed, messages=messages,
_turn_exit_reason=_turn_exit_reason,
_pending_verification_response=_pending_verification_response,
_pending_verification_response_previewed=_pending_verification_response_previewed,
logger=logger,
)
# A non-interrupted turn that fell out of the loop after a tool result, with no
# follow-up assistant text, is the Desktop/TUI "silent stop" (#55316, #54756): the
# composer returns to ready (or keeps spinning) while the durable transcript ends
# at a raw ``tool`` row — the user never learns the turn stopped, and the next user
# message lands as ``tool → user``. Interrupted tails keep
# ``close_interrupted_tool_sequence``; this is the non-interrupt sibling. Mint the
# exit reason, fail the turn, and synthesize the visible close so the tail close in
# ``_persist_step`` persists an assistant row. A turn that already streamed text is
# left alone: ``_recover_final_from_stream`` owns that recovery (#95514).
if (
not final_response
and not interrupted
and messages
and isinstance(messages[-1], dict)
and messages[-1].get("role") == "tool"
and not (getattr(agent, "_current_streamed_assistant_text", "") or "").strip()
):
_turn_exit_reason = "pending_tool_result"
failed = True
final_response = ""
try:
if agent._turn_completion_explainer_enabled():
final_response = (
agent._format_turn_completion_explanation("pending_tool_result", None) or ""
)
except Exception:
final_response = ""
if not final_response:
# The turn-completion explainer opt-out must not reintroduce the silent stop.
final_response = (
"No reply: the turn stopped while a tool result was still pending. "
"Send `continue` to let the model summarize."
)
# Loop exits that are failures in their own right (outer-loop error cap, shutdown, context
# that could not be shrunk) carry the verdict the UI descriptor needs; a bare
# ``turn_exit_reason`` collapsed to code="unknown", retryable=True on every surface.
# Advisory verdicts (``fails_turn=False``) only add the code: ``failed``/``completed`` keep
# the loop's values so cron, kanban and transcript persistence behave as before.
_exit_failure = None if interrupted else exit_reason_failure(_turn_exit_reason)
if _exit_failure is not None and _exit_failure.fails_turn:
failed = True
# Sibling producers (``turn_recovery``, ``codex_runtime``) return ``completed=False`` for an
# interrupted turn; the gateway stream gate and the API run status rely on that contract.
completed = (
final_response is not None
and not failed
and not interrupted
and (api_call_count < agent.max_iterations or str(_turn_exit_reason).startswith("text_response("))
)
_rollback_interrupted_preflight_display(agent, interrupted)
from hermes_cli.observability.shared_metrics_harness import finish_turn
finish_turn(agent, _turn_exit_reason, final_response, interrupted=interrupted, failed=failed)
_cleanup_errors: list[str] = []
# The model has answered (or the loop gave up): a title upgrade held back because it shares a
# self-hosted endpoint with the main request (#117296) may go out now.
from agent.turn_context import start_deferred_title_upgrade
_guarded_cleanup("start_deferred_title_upgrade", lambda: start_deferred_title_upgrade(agent), _cleanup_errors, logger)
# ``user_message`` may be a multimodal list of parts; the trajectory format wants a string.
_guarded_cleanup(
"save_trajectory",
lambda: agent._save_trajectory(messages, _summarize_user_message_for_log(user_message), completed),
_cleanup_errors, logger,
)
_guarded_cleanup(
"cleanup_task_resources", lambda: agent._cleanup_task_resources(effective_task_id),
_cleanup_errors, logger,
)
# Persist only after the transcript tail is shaped and scaffolding removed. Each
# sub-step runs in the same order as the original inline block, and the
# stream-recovered ``final_response`` is rebound the moment it is computed — BEFORE
# the fallible tail-shaping / override / micro-compaction / persist calls — so a
# raise in any of them can't drop text the user already saw (#95514, #8049).
def _persist_step():
nonlocal final_response
_drop_transcript_scaffolding(agent, messages)
final_response, _recovered_from_stream = _recover_final_from_stream(
agent, final_response, interrupted, failed
)
# Recovery paths (stream-recovered / prior-turn text) reach here with a response no
# earlier seam transformed; the normal text turn already did this before its flush and
# gets the recorded outcome back. Either way the tail close below writes the text the
# user will see, never the raw model text (#44239).
if final_response and not interrupted:
_raw = final_response
final_response, _, _ = apply_llm_output_transform(agent, final_response, turn_id=turn_id, logger=logger)
_store_rewrite_on_tail(agent, messages, _raw, final_response)
_close_transcript_tail(agent, messages, final_response, interrupted, _recovered_from_stream)
if not interrupted and not failed and not vetoed:
_micro_compact_after_turn(agent, messages, final_response, logger, effective_task_id)
agent._persist_session(messages, conversation_history)
_guarded_cleanup("persist_session", _persist_step, _cleanup_errors, logger)
# Keep the gateway's separate in-memory history snapshot current even on
# cleanup error, so a later prompt isn't sent with a pre-turn snapshot.
with suppress(Exception):
agent._session_messages = messages
_log_turn_exit(agent, messages, final_response, api_call_count, _turn_exit_reason, interrupted, logger)
# Response transforms apply only to real, uninterrupted responses.
if final_response and not interrupted:
final_response = _append_file_mutation_footer(agent, final_response, logger)
if not interrupted:
final_response = _explain_abnormal_exit(
agent, final_response, _turn_exit_reason, preserved_verification_fallback, logger,
)
_platform = getattr(agent, "platform", None) or ""
_response_transformed = False
_pre_transform_response = None
if final_response and not interrupted:
final_response, _response_transformed, _pre_transform_response = _apply_output_hooks(
agent, final_response, logger, platform=_platform, effective_task_id=effective_task_id,
turn_id=turn_id, original_user_message=original_user_message, messages=messages,
)
# Context engine observation hook: the turn finished with the finalized transcript.
# Fail-open. ``_last_turn_usage`` is the last response's canonical usage dict, or
# ``None`` on turns that never reached a provider response — by contract.
try:
from agent.conversation_loop import _notify_context_engine_turn_complete
_notify_context_engine_turn_complete(
agent, messages, usage=getattr(agent, "_last_turn_usage", None), logger=logger,
turn_id=turn_id, task_id=effective_task_id, api_call_count=api_call_count,
interrupted=interrupted, failed=failed, turn_exit_reason=_turn_exit_reason,
)
except Exception as exc:
logger.warning("on_turn_complete notification failed: %s", exc)
# Surrogate chokepoint: RAW SDK text with a lone UTF-16 surrogate crashes downstream
# consumers (stdout, Telegram ``utf16_len``, JSON); scrub once where it leaves the loop.
# Class-level surrogate chokepoint (#80366, #55143, #55309, #19819): ``final_response`` is often the RAW
# SDK content (``assistant_message.content``), not the sanitized copy stored in history by
# ``build_assistant_message``. Any lone UTF-16 surrogate (U+D800–U+DFFF) in it crashes downstream
# consumers — oneshot stdout writes, Telegram's ``utf16_len`` length check, Signal formatting, JSON
# envelope encodes — on every provider (Ollama, NVIDIA NIM, …). Scrub once here, where model text leaves
# the conversation loop, so every delivery surface receives valid Unicode.
if isinstance(final_response, str):
final_response = _sanitize_surrogates(final_response)
result = {
"final_response": final_response,
"last_reasoning": _last_turn_reasoning(messages),
"messages": messages,
"api_calls": api_call_count,
"completed": completed,
"turn_exit_reason": _turn_exit_reason,
"failed": failed,
"partial": False, # True only when stopped due to invalid tool calls
"interrupted": interrupted,
"response_transformed": _response_transformed,
"pre_transform_response": public_pre_transform(agent, _pre_transform_response),
"response_previewed": getattr(agent, "_response_was_previewed", False),
# The final is byte-for-byte a response this turn already delivered (no footer or
# explanation appended since): it carries no new text for the client to paint.
"response_reused": bool(final_response) and final_response == getattr(agent, "_reused_response_text", None),
"model": agent.model,
# requested_model / served_model: proxy-reported deployment or Hermes' own fallback route.
**result_model_fields(agent),
"provider": agent.provider,
"base_url": agent.base_url,
**{key: getattr(agent, f"session_{key}") for key in _SESSION_TOKEN_KEYS},
# Gateway SessionEntry persists an API reading, never the preflight display seed.
"last_prompt_tokens": (
getattr(agent.context_compressor, "last_real_prompt_tokens", agent.context_compressor.last_prompt_tokens)
if getattr(agent.context_compressor, "last_prompt_tokens", 0) > 0
else getattr(agent.context_compressor, "last_prompt_tokens", 0)
) or 0,
**{key: getattr(agent, f"session_{key}") for key in _SESSION_COST_KEYS},
# Requested service tier, for billing audits (`hermes -z --usage-file`).
"service_tier": (
(getattr(agent, "request_overrides", {}) or {}).get("extra_body") or {}
).get("service_tier"),
"session_id": agent.session_id,
}
if agent._tool_guardrail_halt_decision is not None:
result["guardrail"] = agent._tool_guardrail_halt_decision.to_metadata()
# Persistence failures already set failed=True; also stamp `error` so the gateway
# surfaces status="error" (desktop can toast) instead of a quiet complete frame, plus
# the machine-readable cause 'session_persistence_failed:<locked|compression|...>'.
if failed and str(_turn_exit_reason) == "session_persistence_failed":
from hermes_constants import profile_cli_selector
# Never rebind final_response here: the memory sync and the background-review gate
# below must still see an empty response on a persistence-failed turn.
result["error"] = final_response or (
"session storage could not be written — check the state database "
f"health (`hermes {profile_cli_selector()}doctor`), then send your message again"
)
_cause = getattr(agent, "_last_persistence_error_cause", None)
result["failure_reason"] = "session_persistence_failed:" + (_cause or "unknown")
elif _exit_failure is not None:
if failed:
result["error"] = final_response or str(_turn_exit_reason)
stamp_failure(result, _exit_failure.reason, _exit_failure.retryable)
# Cleanup failures are surfaced, but the response is returned either way (#8049).
if _cleanup_errors:
result["cleanup_errors"] = _cleanup_errors
# A /steer landing after the final assistant turn has no tool batch to drain into;
# hand it back so it becomes the next user turn instead of being lost. A gateway transfer
# (ticketed) stays buffered: the claim's settlement returns it to its own queue place.
_leftover_steer = (getattr(agent, "_drain_leftover_steer", None) or agent._drain_pending_steer)()
if _leftover_steer:
result["pending_steer"] = _leftover_steer
agent._response_was_previewed = False
agent._reused_response_text = None
if interrupted and agent._interrupt_message:
result["interrupt_message"] = agent._interrupt_message
agent.clear_interrupt()
agent._stream_callback = None # don't leak into future calls
# A voice turn's model route ends with the turn: memory sync and the background review
# below (and the next turn) run on the session's main model.
from agent.voice_turn_route import end_voice_turn_route
end_voice_turn_route(agent)
# Skill trigger is checked NOW — based on how many tool iterations THIS turn used.
_should_review_skills = (
agent._skill_nudge_interval > 0
and agent._iters_since_skill >= agent._skill_nudge_interval
and "skill_manage" in agent.valid_tool_names
)
if _should_review_skills:
agent._iters_since_skill = 0
_mirror_turn_to_memory(
agent, original_user_message=original_user_message, final_response=final_response,
interrupted=interrupted, messages=messages, review_memory=_should_review_memory,
review_skills=_should_review_skills, vetoed=vetoed,
)
# Opt-in (compression.prepare_ahead): near the trigger, write the next compaction summary
# between turns. No-op otherwise; never raises. Not after a vetoed turn (no model calls).
from agent.prepared_compaction import maybe_prepare
if not vetoed:
maybe_prepare(agent, messages, origin="turn-end")
# Memory provider on_session_end()/shutdown_all() are NOT called here:
# run_conversation() runs once per message; CLI/gateway own session-end cleanup.
if not getattr(agent, "_persist_disabled", False):
_invoke_hook_safely(
"on_session_end", logger,
session_id=agent.session_id,
task_id=effective_task_id,
turn_id=turn_id,
completed=completed,
failed=failed,
interrupted=interrupted,
turn_exit_reason=_turn_exit_reason,
model=agent.model,
platform=_platform,
)
agent._turn_preflight_display_snapshot = None
agent._turn_received_provider_response = False
return result