feat(tui): rank the slash menu by the skills you actually use
The `/` menu was a flat first-30 slice of the completer's output, and the completer emits every registry command before the first skill. On a 230-skill install that meant a bare `/` filled all 30 rows with commands and offered no skill at all, while `/p` cut off inside the alphabetical skill block — dropping /proving-a-fix-works (471 invocations) and /pr-update (160) but keeping /pretext (2). Spend the limit per kind and rank the skill block by recorded usage (the same .usage.json count Capabilities shows), most-used first and A-Z within a tie. A bare `/` is browsing, so bundled skills that shipped with Hermes and were never opened are dropped as noise; a typed query is a search, and a search that hides a match is broken, so there nothing is pruned and the ranking only reorders. An argument stage keeps the order its own command chose.
This commit is contained in:
@@ -6126,6 +6126,94 @@ def test_complete_slash_reasoning_includes_current_efforts_and_global_scope():
|
||||
assert {"max", "ultra", "--global"} <= values
|
||||
|
||||
|
||||
_SLASH_FILLER_COUNT = 60
|
||||
|
||||
|
||||
def _slash_skill_fixtures(monkeypatch):
|
||||
"""Stub a skill install big enough that a flat cap would truncate it."""
|
||||
filler = {f"/filler-{i:03d}": 0 for i in range(_SLASH_FILLER_COUNT)}
|
||||
usage = {"work": 297, "research": 84, "clean": 12}
|
||||
|
||||
monkeypatch.setattr(
|
||||
server,
|
||||
"_skill_usage_lookup",
|
||||
lambda: (
|
||||
lambda name: usage.get(name, 0),
|
||||
lambda name: "bundled" if name.startswith("unused-") else "local",
|
||||
),
|
||||
)
|
||||
monkeypatch.setattr(
|
||||
"agent.skill_commands.get_skill_commands",
|
||||
lambda: {
|
||||
"/work": {"description": "Fresh worktree"},
|
||||
"/research": {"description": "Look it up"},
|
||||
"/clean": {"description": "Polish the diff"},
|
||||
"/unused-bundled": {"description": "Shipped, never opened"},
|
||||
**{cmd: {"description": "Filler"} for cmd in filler},
|
||||
},
|
||||
)
|
||||
monkeypatch.setattr("agent.skill_bundles.get_skill_bundles", lambda: {})
|
||||
|
||||
|
||||
def _slash_completions(text: str) -> list[dict]:
|
||||
resp = server.handle_request(
|
||||
{"id": "1", "method": "complete.slash", "params": {"text": text}}
|
||||
)
|
||||
return resp["result"]["items"]
|
||||
|
||||
|
||||
def test_complete_slash_offers_skills_alongside_commands(monkeypatch):
|
||||
"""A bare `/` must reach the skills, not just the registry.
|
||||
|
||||
The completer emits every registry command before the first skill, so one
|
||||
flat cap spent every row on commands and no skill was reachable at all.
|
||||
"""
|
||||
_slash_skill_fixtures(monkeypatch)
|
||||
|
||||
kinds = {item["kind"] for item in _slash_completions("/")}
|
||||
|
||||
assert kinds == {"command", "skill"}
|
||||
|
||||
|
||||
def test_complete_slash_ranks_skills_by_recorded_usage(monkeypatch):
|
||||
"""The skills someone actually invokes lead the ones they never opened."""
|
||||
_slash_skill_fixtures(monkeypatch)
|
||||
|
||||
skills = [
|
||||
item["text"].strip() for item in _slash_completions("/") if item["kind"] == "skill"
|
||||
]
|
||||
|
||||
assert skills[:3] == ["work", "research", "clean"]
|
||||
|
||||
|
||||
def test_complete_slash_prunes_unused_builtins_only_while_browsing(monkeypatch):
|
||||
"""A bare `/` is browsing and may prune; a typed query is a search.
|
||||
|
||||
A search that hides a match is broken, so the never-opened bundled skill
|
||||
disappears from `/` and comes straight back the moment it is typed for.
|
||||
"""
|
||||
_slash_skill_fixtures(monkeypatch)
|
||||
|
||||
browsing = {item["text"].strip() for item in _slash_completions("/")}
|
||||
searching = {item["text"].strip() for item in _slash_completions("/unused")}
|
||||
|
||||
assert "unused-bundled" not in browsing
|
||||
assert "unused-bundled" in searching
|
||||
|
||||
|
||||
def test_complete_slash_leaves_argument_stages_alone(monkeypatch):
|
||||
"""Ranking applies to the command token, never to a command's own args.
|
||||
|
||||
`/details c` completes that command's modes; a skill named /clean also
|
||||
starts with a `c` and must not be offered as one of them.
|
||||
"""
|
||||
_slash_skill_fixtures(monkeypatch)
|
||||
|
||||
items = _slash_completions("/details c")
|
||||
|
||||
assert [item["text"] for item in items] == ["collapsed", "cycle"]
|
||||
|
||||
|
||||
def test_config_set_reasoning_updates_live_session_and_agent(tmp_path, monkeypatch):
|
||||
monkeypatch.setattr(server, "_hermes_home", tmp_path)
|
||||
(tmp_path / "config.yaml").write_text("agent:\n reasoning_effort: medium\n", encoding="utf-8")
|
||||
|
||||
@@ -259,7 +259,20 @@ def _(rid, params: dict) -> dict:
|
||||
),
|
||||
}
|
||||
for c in completer.get_completions(doc, None)
|
||||
][:30]
|
||||
]
|
||||
|
||||
# Rank and bound the list (see _rank_slash_completions) while a
|
||||
# `/token` is under the cursor — the one stage skills are offered at.
|
||||
# An argument stage (`/personality `, `/details c`) keeps the order
|
||||
# its own command chose.
|
||||
if text.rsplit(" ", 1)[-1].startswith("/"):
|
||||
usage, origin_of = _skill_usage_lookup()
|
||||
items = _rank_slash_completions(
|
||||
items, usage, origin_of, browsing=text == "/"
|
||||
)
|
||||
else:
|
||||
items = items[:_SLASH_COMPLETION_LIMIT]
|
||||
|
||||
text_lower = text.lower()
|
||||
extras = [
|
||||
{
|
||||
|
||||
@@ -11395,6 +11395,55 @@ def _skill_usage_lookup():
|
||||
return usage, origin
|
||||
|
||||
|
||||
_SLASH_COMPLETION_LIMIT = 30
|
||||
|
||||
|
||||
def _rank_slash_completions(
|
||||
items: list[dict],
|
||||
usage,
|
||||
origin_of,
|
||||
*,
|
||||
browsing: bool,
|
||||
) -> list[dict]:
|
||||
"""Rank and bound slash completions the way the menu should read.
|
||||
|
||||
``usage``/``origin_of`` are the callables :func:`_skill_usage_lookup`
|
||||
returns. Registry commands keep their existing order — only the skill
|
||||
block is reordered, most-used first and A-Z within a tie, so the handful
|
||||
of skills someone invokes daily lead the ones that shipped with Hermes
|
||||
and were never opened.
|
||||
|
||||
The limit is spent PER KIND rather than on one flat truncation. A flat
|
||||
cut is positional, not editorial: the completer emits every registry
|
||||
command before the first skill, so on a 230-skill install a bare ``/``
|
||||
hit the cap while still inside the command block and offered no skill at
|
||||
all, and ``/p`` dropped ``/proving-a-fix-works`` (471 uses) while keeping
|
||||
``/pretext`` (2).
|
||||
|
||||
``browsing`` separates the two things a slash means. A bare ``/`` is
|
||||
BROWSING, so bundled skills with no recorded activity are dropped as
|
||||
noise. A typed query is SEARCHING, and a search that hides a match is
|
||||
broken — there nothing is pruned, the ranking only reorders.
|
||||
"""
|
||||
|
||||
def name_of(item: dict) -> str:
|
||||
return str(item.get("text", "")).strip().lstrip("/").lower()
|
||||
|
||||
commands = [item for item in items if item.get("kind") != "skill"]
|
||||
skills = [item for item in items if item.get("kind") == "skill"]
|
||||
|
||||
if browsing:
|
||||
skills = [
|
||||
item
|
||||
for item in skills
|
||||
if origin_of(name_of(item)) != "bundled" or usage(name_of(item)) > 0
|
||||
]
|
||||
|
||||
skills.sort(key=lambda item: (-usage(name_of(item)), name_of(item)))
|
||||
|
||||
return commands[:_SLASH_COMPLETION_LIMIT] + skills[:_SLASH_COMPLETION_LIMIT]
|
||||
|
||||
|
||||
def _cli_exec_blocked(argv: list[str]) -> str | None:
|
||||
"""Return user hint if this argv must not run headless in the gateway process."""
|
||||
if not argv:
|
||||
|
||||
Reference in New Issue
Block a user