71cb99136a
`_fresh_cache` cleared `bot_mode_probe._cached`, which `capability_fingerprint` is documented as deliberately never touching; the `home` fixture wrote a hermes-bots managed `profile.yaml` that no assertion reads (every test compares digests before and after edits to the skills tree only). PROOF: `pytest tests/tools/test_capability_fingerprint_skill_scope.py` → 2 passed with both removed; `ruff check` clean; footguns clean.
60 lines
2.4 KiB
Python
60 lines
2.4 KiB
Python
"""The capability epoch tracks the skills the model can actually invoke.
|
|
|
|
``capability_fingerprint`` walks the profile's skills tree, and a change to its digest rebuilds the
|
|
stored Bot Chat system prompt — re-prefilling the whole prompt cache behind it. It must therefore
|
|
move for a real capability change and stay still for anything that is not one.
|
|
|
|
A raw ``**/SKILL.md`` glob counted files under the dirs every other reader of this tree prunes
|
|
(``agent.skill_utils.EXCLUDED_SKILL_DIRS``: ``.archive``, ``.curator_backups``, ``node_modules`` …)
|
|
and each skill's support dirs — files ``skills_list``/``skill_view`` never offer, ``skill_count``
|
|
never counts, and the prompt's own skills index never lists. So archiving a skill, or the curator
|
|
writing a backup, flipped the epoch and rebuilt a prompt whose skills section had not changed.
|
|
"""
|
|
from __future__ import annotations
|
|
|
|
import pytest
|
|
|
|
from tools import bot_mode_probe
|
|
|
|
|
|
@pytest.fixture
|
|
def home(tmp_path):
|
|
h = tmp_path / ".hermes"
|
|
h.mkdir()
|
|
return h
|
|
|
|
|
|
def _install(home, relpath: str, name: str = "s") -> None:
|
|
d = home / "skills" / relpath
|
|
d.mkdir(parents=True, exist_ok=True)
|
|
(d / "SKILL.md").write_text(f"---\nname: {name}\n---\n", encoding="utf-8")
|
|
|
|
|
|
def test_installing_a_real_skill_still_moves_the_epoch(home):
|
|
before = bot_mode_probe.capability_fingerprint(home)
|
|
|
|
_install(home, "web/scraping", "scraping")
|
|
|
|
assert bot_mode_probe.capability_fingerprint(home) != before
|
|
|
|
|
|
def test_archiving_a_skill_does_not_move_the_epoch(home):
|
|
_install(home, "web/scraping", "scraping")
|
|
_install(home, "web/keep", "keep")
|
|
before = bot_mode_probe.capability_fingerprint(home)
|
|
|
|
# What `skills archive` leaves behind: the package moved under `.archive/`.
|
|
src = home / "skills" / "web" / "scraping"
|
|
dst = home / "skills" / ".archive" / "scraping"
|
|
dst.parent.mkdir(parents=True, exist_ok=True)
|
|
src.rename(dst)
|
|
after_archive = bot_mode_probe.capability_fingerprint(home)
|
|
|
|
# It DID leave the invocable set, so the epoch moves once...
|
|
assert after_archive != before
|
|
# ...and the archived copy sitting there is not itself a capability: re-running is stable,
|
|
# and adding more archived packages never moves it again.
|
|
assert bot_mode_probe.capability_fingerprint(home) == after_archive
|
|
_install(home, ".archive/another-old-skill", "another")
|
|
assert bot_mode_probe.capability_fingerprint(home) == after_archive
|