mirror of
https://github.com/NousResearch/hermes-agent.git
synced 2026-11-17 04:11:35 +00:00
An explicit SESSION_ID whose stored prompt is already NULL used to print "No degraded stored system prompts found.", which reads like a scan result. Say the row already has no stored prompt and there is nothing to clear. The SESSION_ID override assertions (memory-only clear, empty-row message, unknown id -> exit 1) move into the memory-only scan test so the stack keeps two invariant tests.
104 lines
4.3 KiB
Python
104 lines
4.3 KiB
Python
"""Behavior contracts for hermes sessions repair-prompts (#122822)."""
|
|
|
|
from __future__ import annotations
|
|
|
|
import json
|
|
from types import SimpleNamespace
|
|
|
|
import pytest
|
|
|
|
from agent.prompt_builder import SKILL_SAFETY_HEADING
|
|
from hermes_cli import sessions_cmd
|
|
from hermes_cli.sessions_cmd import _cmd_repair_prompts
|
|
from hermes_state import SessionDB
|
|
|
|
HEALTHY = (
|
|
"You are Hermes.\n"
|
|
"<available_skills>\n dogfood: exploratory QA of web apps\n</available_skills>\n"
|
|
f"{SKILL_SAFETY_HEADING}\nReload [SKILL_PRUNED] placeholders with skill_view.\n"
|
|
)
|
|
DEGRADED = "You are Hermes.\n(reduced maintenance build without the skills index)\n"
|
|
|
|
|
|
def _tool(name: str) -> dict:
|
|
return {
|
|
"type": "function",
|
|
"function": {
|
|
"name": name,
|
|
"description": f"{name} test tool",
|
|
"parameters": {"type": "object", "properties": {}},
|
|
},
|
|
}
|
|
|
|
|
|
def _pin(*names: str) -> dict:
|
|
return {"version": "test-sha", "tools": [_tool(name) for name in names]}
|
|
|
|
|
|
@pytest.fixture()
|
|
def db(tmp_path):
|
|
session_db = SessionDB(db_path=tmp_path / "state.db")
|
|
yield session_db
|
|
session_db.close()
|
|
|
|
|
|
def _args(session_id=None, apply=False, json_out=False):
|
|
return SimpleNamespace(session_id=session_id, apply=apply, json=json_out)
|
|
|
|
|
|
def _apply_json(db, monkeypatch, capsys):
|
|
monkeypatch.setattr(
|
|
sessions_cmd, "_confirm_prompt",
|
|
lambda _prompt: pytest.fail("--apply --json must not prompt"),
|
|
)
|
|
assert _cmd_repair_prompts(db, _args(apply=True, json_out=True)) == 0
|
|
return json.loads(capsys.readouterr().out)
|
|
|
|
|
|
def test_unpinned_row_is_unverifiable_and_apply_json_clears_only_verified(db, monkeypatch, capsys):
|
|
unverifiable = db.create_session("old-unpinned", "telegram", system_prompt=DEGRADED)
|
|
verified = db.create_session("verified", "telegram", system_prompt=DEGRADED)
|
|
db.update_session_tool_names(verified, _pin("skills_list", "skill_manage"))
|
|
healthy = db.create_session("healthy-1", "telegram", system_prompt=HEALTHY)
|
|
db.update_session_tool_names(healthy, _pin("skills_list", "skill_manage"))
|
|
|
|
payload = _apply_json(db, monkeypatch, capsys)
|
|
|
|
assert payload["apply"] is True
|
|
assert payload["cleared"] == ["verified"]
|
|
assert [row["id"] for row in payload["unverifiable"]] == ["old-unpinned"]
|
|
assert db.get_session(unverifiable)["system_prompt"] == DEGRADED
|
|
assert not (db.get_session(verified)["system_prompt"] or "")
|
|
assert db.get_session(verified)["tool_names"] # the skill_manage pin survives the clear
|
|
assert db.get_session(healthy)["system_prompt"] == HEALTHY
|
|
|
|
|
|
def test_scan_leaves_reduced_memory_only_zero_skills_rows_and_session_id_overrides(db, monkeypatch, capsys):
|
|
reduced = db.create_session("reduced-1", "telegram", system_prompt=DEGRADED)
|
|
db.update_session_tool_names(reduced, _pin("todo", "web_search"))
|
|
memory = db.create_session("memory-only-1", "telegram", system_prompt=DEGRADED)
|
|
db.update_session_tool_names(memory, _pin("memory"))
|
|
# Zero skills installed + no skill_manage: the real builder emits neither skills marker.
|
|
zero_skills = "You are Hermes.\nNo skills are installed.\n"
|
|
readonly = db.create_session("zero-skills-readonly", "telegram", system_prompt=zero_skills)
|
|
db.update_session_tool_names(readonly, _pin("terminal", "skills_list", "skill_view"))
|
|
|
|
payload = _apply_json(db, monkeypatch, capsys)
|
|
|
|
assert payload["cleared"] == []
|
|
assert [row["id"] for row in payload["unverifiable"]] == ["memory-only-1"]
|
|
assert db.get_session(reduced)["system_prompt"] == DEGRADED
|
|
assert db.get_session(memory)["system_prompt"] == DEGRADED
|
|
assert db.get_session(memory)["tool_names"]
|
|
assert db.get_session(readonly)["system_prompt"] == zero_skills
|
|
|
|
# An explicit SESSION_ID is the operator override for the unverifiable memory-only row.
|
|
monkeypatch.setattr(sessions_cmd, "_confirm_prompt", lambda _prompt: True)
|
|
assert _cmd_repair_prompts(db, _args(session_id="memory-only", apply=True)) == 0
|
|
assert not (db.get_session(memory)["system_prompt"] or "")
|
|
|
|
capsys.readouterr()
|
|
assert _cmd_repair_prompts(db, _args(session_id="memory-only", apply=True)) == 0
|
|
assert "already has no stored prompt" in capsys.readouterr().out
|
|
assert _cmd_repair_prompts(db, _args(session_id="no-such-session", apply=True)) == 1
|
|
assert "No session matches" in capsys.readouterr().out
|