mirror of
https://github.com/NousResearch/hermes-agent.git
synced 2026-07-24 16:54:43 +00:00
cbf5b05c70
A follow-up sent while the model is still generating previously ended the turn: Hermes kept only the visible partial text (reasoning was display-only), cleared the loop, and replayed the message as a fresh next turn. If the correction referred to something that only appeared in the thinking stream, the model no longer had that context. Add `AIAgent.redirect(text)`: a corrective interrupt distinct from a hard stop. It cancels only the in-flight model request (not tool workers or child agents), stashes the correction under a lock shared with `interrupt()` so a concurrent `/stop` always wins, and lets the loop rebuild the same logical iteration. `_apply_active_turn_redirect()` checkpoints the reasoning that was actually shown to the user plus any visible partial text as an ordinary assistant message, then appends the correction as a real user turn — never replaying incomplete signed/encrypted provider reasoning, and keeping strict role alternation and prompt-cache stability intact. During tool execution it degrades to `steer()` so a running tool finishes at a safe boundary. `_fire_reasoning_delta` now only records reasoning that a display callback actually consumed, so `show_reasoning: false` never leaks hidden provider thinking into the persisted transcript.
68 lines
2.3 KiB
Python
68 lines
2.3 KiB
Python
"""Unit tests for TurnRetryState (god-file Phase 1b).
|
|
|
|
The dataclass holds the inner-retry-loop's one-shot recovery guards + restart
|
|
signals. These tests pin its shape and default semantics — the behavioral
|
|
guarantee for the loop itself is the existing recovery-branch tests in
|
|
tests/run_agent/ which now exercise these fields via `_retry.<flag>`.
|
|
"""
|
|
|
|
from __future__ import annotations
|
|
|
|
from dataclasses import fields
|
|
|
|
from agent.turn_retry_state import TurnRetryState
|
|
|
|
|
|
EXPECTED_FIELDS = {
|
|
"codex_auth_retry_attempted",
|
|
"anthropic_auth_retry_attempted",
|
|
"nous_auth_retry_attempted",
|
|
"nous_paid_entitlement_refresh_attempted",
|
|
"copilot_auth_retry_attempted",
|
|
"vertex_auth_retry_attempted",
|
|
"thinking_sig_retry_attempted",
|
|
"invalid_encrypted_content_retry_attempted",
|
|
"image_shrink_retry_attempted",
|
|
"multimodal_tool_content_retry_attempted",
|
|
"oauth_1m_beta_retry_attempted",
|
|
"llama_cpp_grammar_retry_attempted",
|
|
"primary_recovery_attempted",
|
|
"has_retried_429",
|
|
"auth_failover_attempted",
|
|
"restart_with_compressed_messages",
|
|
"restart_with_length_continuation",
|
|
"restart_with_rebuilt_messages",
|
|
"restart_with_redirected_messages",
|
|
}
|
|
|
|
|
|
def test_all_guards_default_false():
|
|
s = TurnRetryState()
|
|
for name, value in s:
|
|
assert value is False, f"{name} should default to False"
|
|
|
|
|
|
def test_field_set_matches_contract():
|
|
names = {f.name for f in fields(TurnRetryState)}
|
|
assert names == EXPECTED_FIELDS, (
|
|
f"unexpected drift: missing={EXPECTED_FIELDS - names} extra={names - EXPECTED_FIELDS}"
|
|
)
|
|
|
|
|
|
def test_loop_control_vars_are_not_on_state():
|
|
# retry_count / max_retries / max_compression_attempts stay as loop locals,
|
|
# NOT on the state object (they are while-mechanics, not recovery bookkeeping).
|
|
names = {f.name for f in fields(TurnRetryState)}
|
|
for loop_local in ("retry_count", "max_retries", "max_compression_attempts"):
|
|
assert loop_local not in names
|
|
|
|
|
|
def test_guards_are_independently_mutable():
|
|
s = TurnRetryState()
|
|
s.codex_auth_retry_attempted = True
|
|
s.restart_with_compressed_messages = True
|
|
assert s.codex_auth_retry_attempted is True
|
|
assert s.restart_with_compressed_messages is True
|
|
# untouched guards stay False
|
|
assert s.has_retried_429 is False
|
|
assert s.anthropic_auth_retry_attempted is False
|